01 What It Is
GPT-6.1 Sol is a reasoning model in OpenAI's GPT-6 family, the middle tier between Astra (top) and Luna (low cost). It launched at DevDay on September 29, 2026 — seven days after GPT-6 Sol — and replaces it. OpenAI's model catalog now calls the old gpt-6-sol the "previous generation workhorse model," and Artificial Analysis noted the swap came "after just 7 days."
The API ID is gpt-6.1-sol. The launch claim, quoted by TechCrunch, is that it "delivers nearly the same level of intelligence as GPT-6 Astra for agentic coding, computer use, and professional work." It is now the default model in Codex and is available in the API, ChatGPT Work, GitHub Copilot and OpenRouter.
02 The Specs
It takes text and image input and returns text — no audio or video in or out. Reasoning effort runs low, medium (default), high, xhigh and max. The context window is 1,050,000 tokens with a 128,000-token output cap, the same as the rest of the GPT-6 line. The knowledge cutoff is April 30, 2026, ten days later than GPT-6 Sol's April 20.
- API model ID —
gpt-6.1-sol - Context / output — 1,050,000 in, 128,000 out
- Reasoning effort — low, medium (default), high, xhigh, max
- Built-in tools — web search, file search, code interpreter, hosted shell, computer use, MCP, tool search
- Endpoints — Chat Completions, Responses, Batch
none reasoning setting — OpenAI's docs say "none and minimal reasoning efforts are not supported." If your GPT-6 Sol code sets effort: "none" for fast, cheap calls, switch it to low before you move to this model, or the request fails. For tool calling, build on the Responses API; Chat Completions is supported but without tool calling.
03 The Benchmarks
Every number here is from OpenAI's announcement and the accompanying system-card addendum, with independent reporting where noted. Different benchmarks measure different things, so read them as coverage, not a single score.
| Benchmark | Tests | GPT-6.1 Sol | Reference |
|---|---|---|---|
| DeepSWE v1.1 | Real-repo software engineering | 75.2% (high) | Astra ~74.8%, Sol 68.8%, Sonnet 5.5 71.0% |
| OSWorld 2.0 | Long-horizon desktop computer use | 71.4% (max) | Sol 64.4%, Astra 73.5% |
| GDP.pdf | Complex professional documents | 32.0% | Astra 32.2%, Opus 5.5 (fallbacks) 28.8% |
| AutomationBench 1.0.6 | Multi-step SaaS workflows | 35.4% (medium) | ~2.2pp above Opus 5.5 (medium) |
| Terminal-Bench Science 0.1 | Scientific terminal work | ~2x GPT-6 Sol | Astra hits the 68.1% ceiling |
| Factual errors (low effort) | Hallucination rate | 7.7% | GPT-6 Sol 11.4% |
source: openai gpt-6.1 sol announcement & system-card addendum, sep 29 2026 · vellum + codersera analysis · the new stack
The headline is DeepSWE: 75.2 at high effort, matching Astra's ~74.8 while costing roughly $1.50 per task against Astra's ~$7.70. On computer use it lands within 2.1 points of Astra. On professional documents it beats Claude Opus 5.5 with fallbacks and sits a tenth of a point under Astra. On factuality, hallucinations at low reasoning effort dropped 32% over the old Sol, and its error rate stays within 1.9 points of Astra across every tier.
04 The Cost Math
This is where the model earns its tier. Standard pricing is $2 input and $10 output per million tokens, and cached input fell from $0.20 to $0.10 — a 95% discount off standard input.
| Mode | GPT-6.1 Sol / 1M | Note |
|---|---|---|
| Standard | $2 in / $10 out | cached input $0.10 |
| Long context (>272K in) | $4 in / $15 out | 2x input, 1.5x output |
| Batch / Flex | $1 in / $5 out | half standard |
| Fast | $4 in / $20 out | 2x standard |
The 272K-token threshold is the one to watch: above it, the whole request bills at 2x input and 1.5x output. For an agent that re-reads a 200K-token repository every step, the new cache price drops per-step input cost from $0.40 to $0.02.
05 The Ceiling, and the Cancellation
GPT-6.1 Sol is not the ceiling. On the scientific terminal benchmark, GPT-6 Astra still defines the top at 68.1% — for novel math proofs and lab-scale analysis, Astra remains the recommended model. Sol is the volume model: it covers the 72% to 75% DeepSWE band between $0.50 and $1.50 per task, where Astra needs $2.00 to $7.70.
06 How It Fits the GPT-6 Line
GPT-6.1 Sol is an updated middle tier, not a new one. It replaces GPT-6 Sol, which had run for a week. The GPT-6 generation, all in September 2026, runs on three tiers:
| GPT-6 tier | Released | $ / 1M in → out | Cached in | Positioning |
|---|---|---|---|---|
| GPT-6 Astra | Sep 3, 2026 | $10 → $50 | $1.00 | Flagship; hardest end-to-end work |
| GPT-6 Sol | Sep 22, 2026 | $2 → $10 | $0.20 | Previous workhorse (replaced) |
| GPT-6 Luna | Sep 22, 2026 | $0.10 → $0.50 | $0.01 | Low-cost, high-volume work |
| GPT-6.1 Sol | Sep 29, 2026 | $2 → $10 | $0.10 | Near-Astra coding at mid-tier price |
OpenAI also announced an Ultrafast tier — up to 8x faster token generation in Codex, up to 6x in the API, around 300 tokens per second, at 6x the standard price. It is live for GPT-6 Astra today; a GPT-6.1 Sol version was "coming soon" at DevDay.
07 The Bottom Line
- Everyday autonomous coding, computer use and documents: GPT-6.1 Sol is the default — near-Astra results at a fifth the per-task cost.
- Long agent loops: the $0.10 cached-input price is the real savings; a 200K-token repo re-read costs $0.02 per step.
- When you need the ceiling: GPT-6 Astra. It still owns the hardest reasoning and the 68.1% scientific top.
- High-volume, low-stakes jobs: GPT-6 Luna at $0.10/$0.50.
This article was drafted with AI assistance and reviewed by a human editor.